深度学习训练方法介绍

本节介绍如何使用keras做深度学习训练。摘自tensorflow.keras官方教程，是手写数字识别的升级版，给定10种衣服类型进行图像识别。原网址见https://www.tensorflow.org/tutorials/keras/classification

直接上代码，配上注释。

# 0. load libs
import tensorflow as tf
from tensorflow import keras
import numpy as np
#import matplotlib.pyplot as plt

# 1. load dataset and preprocess
fashion_mnist = keras.datasets.fashion_mnist                #Fashion MNIST是一个预定义好的训练例子，可以直接下载读取
(train_images, train_labels), (test_images, test_labels) = fashion_mnist.load_data()        #已经区分好了60000训练10000验证
#图片是28x28像素，灰度表示，每个像素的值是0-255，标签是0-9，代表9种不同的衣服种类
#train_images.shape   ->      (60000, 28, 28)
#len(train_labels)      ->     60000
#print(train_labels)

train_images = train_images / 255.0    #Scale values to a range of 0 to 1 before feeding them to the neural network model
test_images = test_images / 255.0
# 2. define model
model = keras.Sequential([                                      #使用构造函数一次性建立神经网络结构
    keras.layers.Flatten(input_shape=(28, 28)),        #Flatten较为特殊，没有参数，起到的作用仅仅是将28x28矩阵转成一维数组(flatten含义)，即将图像输入转为合理格式
    keras.layers.Dense(128, activation='relu'),
    keras.layers.Dense(10, activation='softmax')       #softmax用于多分类过程中，将多个神经元的输出映射到（0,1）区间内，即最后结果的概率
])
# 3. compile model
model.compile(optimizer='adam', loss='sparse_categorical_crossentropy', metrics=['accuracy'])           #loss - 计算预测值与真值的多类交叉熵(输入值为二值矩阵,而不是向量)
# 4. train model
model.fit(train_images, train_labels, epochs=10)            #The model learns to associate images and labels
# 5. evaluate model
test_loss, test_acc = model.evaluate(test_images, test_labels)
print('\nTest accuracy:', test_acc)
# 6. make predictions - single
img = test_images[678]
imgA = (np.expand_dims(img,0))      #model做预测都是batch，即使只有一个元素也要变成list
predictions_single = model.predict(imgA)
print(predictions_single)                   #会给出10个标签各自的概率
#print("Prediction: " + str(np.argmax(predictions_single[0]) + "Correct answer: " + str(test_labels[678])))
print(np.argmax(predictions_single[0]))
print(test_labels[678])